侯位昭 齐幸辉 宋凯磊 韩志卓 司佳 刘勇

摘  要： 警情的下发效率直接关系到公安民警的出警质量，从而影响公安机关驾驭治安局势的能力和水平。在以往的警情下发中，一般只考虑地址位置的经纬度信息，这就导致在管辖范围界限模糊、经纬度定位不准确时，仍需手动下发警情，容易造成人力、物力的损失以及区域安全指数的降低。为了解决这个问题，文中创新性地将区域划分问题转化为分类问题，并综合考虑经纬度地理编码信息以及中文地名语义信息，提出基于FastText和WKNN的融合地址位置和文本相似性的警情自动下发方法。该方法融合了由FastText得到的地址词向量和根据地址解析服务得到的经纬度信息，将二者组成的地址多元要素作为加权KNN（WKNN）模型的输入来训练分类器。同时，WKNN采用sigmoid函数自适应地权衡在不同经纬度解析精度下地址位置坐标与词向量相似性的权重，提高了模型的鲁棒性。以某市历史警情下发数据为依托，实验结果显示警情下发准确率在91%以上，验证了该模型在某市经纬度不准确、新地址冷启动等警情下发场景中的有效性及高效性。

关键词： 警情自动下发; 融合模型; 信息融合; 权重权衡; 分类器训练; 模型验证

中图分类号： TN911.1?34; TP181; TP391               文献标识码： A              文章编号： 1004?373X（2020）13?0073?08

Intelligent police alarm issuing based on FastText and WKNN fusion model

HOU Weizhao1， QI Xinghui1， SONG Kailei1， HAN Zhizhuo1， SI Jia1， LIU Yong2

（1. The 54th Research Institute， China Electronics Technology Group Corporation， Shijiazhuang 050081， China;

2. Institute of Information Engineering， Chinese Academy of Sciences， Beijing 100093， China）

Abstract： The issuing efficiency of the police alarm is directly related to the response effect of public security police， which accordingly influence the ability and level of the public security organs to control the security situation. In the past police alarm issuing， only the latitude and longitude information of an address location has been taken into account. Therefore， it has led to manual police alarm issuing when jurisdiction boundaries are obscure， and latitude and longitude positioning is inaccurate， which is prone to cause the loss of manpower and material resources， and the reduction of regional safety index. In view of the above， the regional division pattern is innovatively transformed into the classification one. In addition， the latitude and longitude geocoding information and the semantic information of Chinese geographical names are taken into account comprehensively， and the FastText and WKNN （weighted K?nearest neighbor） based automatic police alarm issuing method is proposed， which fuses the address word vector got by FastText and the latitude and longitude information got by address resolution service. The address multivariate elements formed by the two are used as the input of weighted KNN （WKNN） model to train the classifier. At the same time， the WKNN with sigmoid function is used to adaptively weigh the weights of address location coordinates and the word vector similarity at different latitude and longitude resolution accuracy， which improves the robustness of the model. The experimental results （which take the historical data of police alarm issuing of a certain city as an example） show that the accuracy of the police alarm issuing is higher than 91%. It verifies the validity and efficiency of the model in situations of inaccurate longitude and latitude and cold start of new address.

Keywords： automatic police alarm issuing; fusion model; signal fusion; weight balance; classifier training; model verification

0  引  言

随着人工智能时代的到来，如何探索一种准确而实用的警情自动下发机制，缩短警情出现到出警的时间周期，提高公安民警的出警质量和效率[1]，成为公安行业内亟待解决的问题。……